introspection-sdk 0.4.4

Python SDK for Introspection — continuously improve your AI systems with production feedback and frontier practices

Build frontier AI systems that self-improve.

Introspection continuously improves your AI systems with production feedback and frontier practices. This is the Python SDK.

Install

uv add introspection-sdk
# or
pip install introspection-sdk

The default install is REST-only — no OpenTelemetry pulled in. Add the [otel] extra to enable analytics events and trace export:

pip install 'introspection-sdk[otel]'

Optional extras

Per-framework convenience installs (init() auto-detects frameworks however they were installed — these are just one-command setup):

pip install 'introspection-sdk[anthropic]'      # Anthropic SDK
pip install 'introspection-sdk[gemini]'         # Google Gemini (google-genai)
pip install 'introspection-sdk[openai-agents]'  # OpenAI Agents SDK
pip install 'introspection-sdk[claude-agent]'   # Claude Agent SDK
pip install 'introspection-sdk[langchain]'      # LangChain / LangGraph
pip install 'introspection-sdk[logfire]'        # Logfire
pip install 'introspection-sdk[all]'            # Everything above

Three independent surfaces

The Python SDK exposes three surfaces you can adopt independently:

1. Introspection API (runtimes, tasks, files) with IntrospectionClient — the main Introspection API. Zero OpenTelemetry imports. Always available.
2. Analytics events (track, feedback, identify) with IntrospectionLogs — OTel logs exporter with baggage helpers. Owns its own LoggerProvider. Lives at introspection_sdk.IntrospectionLogs. Requires the [otel] extra.
3. Traces (span processors + instrumentors) with IntrospectionSpanProcessor and friends — IntrospectionTracingProcessor, ClaudeTracingProcessor, the LangChain callback handler, AnthropicInstrumentor, GeminiInstrumentor. Plus the introspection_sdk.init() convenience that auto-wires every supported framework. All under introspection_sdk.otel (or the dedicated introspection_sdk.integrations.langchain subpath for the LangChain handler). Requires the [otel] extra.

1. Introspection API (runtimes, tasks, files) with IntrospectionClient

The main Introspection API surface. No OTel packages required — install just the SDK:

pip install introspection-sdk

from introspection_sdk import IntrospectionClient

client = IntrospectionClient()  # token from INTROSPECTION_TOKEN

runner = client.runtimes("customer-agent").run()

run = runner.tasks.start(prompt="Say hello in one sentence.")

for event in run.stream():
    print(f"[{event.event}] {event.data}")

runner.close()
client.shutdown()

See examples/api/runtimes.py for an end-to-end walkthrough.

2. Analytics events (track, feedback, identify) with IntrospectionLogs

Install the SDK with the [otel] extra:

pip install 'introspection-sdk[otel]'

from introspection_sdk import IntrospectionLogs

logs = IntrospectionLogs(
    token="intro_xxx",        # or env: INTROSPECTION_TOKEN
    service_name="my-service",
    base_url="https://otel.introspection.dev",  # or env: INTROSPECTION_BASE_OTEL_URL
    project_id="proj_…",      # or env: INTROSPECTION_PROJECT_ID — optional
)

with logs.identify("user_123", traits={"plan": "pro"}):
    with logs.conversation("conv_456", previous_response_id="msg_123"):
        logs.feedback("thumbs_up", comments="Great response!")

logs.track("checkout_completed", {"amount": 42})
logs.shutdown()

Methods

Method	Description
track(event, properties=)	Track any user action
feedback(type, **kwargs)	Track feedback on AI responses
identify(user_id, traits=)	Associate a user with traits (context manager)
flush(timeout_ms=30000)	Flush pending events
shutdown()	Shutdown and flush

Context managers

Method	Description
conversation(id?, previous_response_id?)	Set conversation context
with_user_id(id)	Set user context
with_agent(name, id?)	Set agent context
with_anonymous_id(id)	Set anonymous ID
with_baggage(**values)	Set arbitrary baggage values

3. Traces (span processors + instrumentors) with IntrospectionSpanProcessor

Install the SDK with the [otel] extra plus your framework extras of choice (or [all]):

pip install 'introspection-sdk[otel,anthropic,gemini,openai-agents,claude-agent,langchain]'

Auto-wired via init() — recommended

introspection.init() detects every supported LLM framework you have installed and wires them all into a single trace pipeline:

import introspection_sdk as introspection

introspection.init()  # token from INTROSPECTION_TOKEN

# ...use Anthropic, Gemini, OpenAI Agents, Claude Agent, Logfire as usual —
# their calls are now traced automatically.

Auto-detected frameworks: Anthropic SDK, Google Gemini (google-genai), OpenAI Agents SDK, Claude Agent SDK, Logfire / OpenInference (configure Logfire before init()), and LangChain / LangGraph (attach get_handler() — see below).

LangChain callbacks are per-invoke, so init() prepares the handler and you attach it:

import introspection_sdk as introspection
from introspection_sdk.integrations.langchain import get_handler

introspection.init()
response = model.invoke("Hello!", config={"callbacks": [get_handler()]})

After init(), the module-level introspection.track() / introspection.feedback() / introspection.identify() shortcuts proxy to the global IntrospectionLogs.

Manual / advanced setup

init() is the recommended entry point, but the individual processors and instrumentors remain fully supported for custom wiring (sharing a TracerProvider, dual-export, testing). See docs/advanced.md for opting out of auto-discovery, passing your own provider, standalone processor construction, and testing with an in-memory exporter.

See examples/ for complete integration patterns including dual-export with Arize, Langfuse, Braintrust, and LangSmith.

Environment variables

# Introspection API (IntrospectionClient)
export INTROSPECTION_TOKEN="intro_xxx"
export INTROSPECTION_BASE_API_URL="https://api.introspection.dev"   # optional
export INTROSPECTION_PROJECT_ID="proj_…"                            # optional

# OTel (IntrospectionLogs + span processors + instrumentors)
export INTROSPECTION_BASE_OTEL_URL="https://otel.introspection.dev" # optional
export INTROSPECTION_SERVICE_NAME="my-service"                      # optional

Documentation

Full documentation is available at docs.introspection.dev.

License

Apache-2.0